Experimental Design
Research design:
This trial employs a cluster-randomized design, where the unit of analysis is a household, and the cluster is a village. The intervention (henceforth the “treatment”) is the delivery of a rainwater cistern coupled with rural extension services and a cash grant of either R$ 1,500 or R$ 3,000 to the household. The choice between the two cash grant levels depends on the household’s poverty profile and the intervention’s budget.
Selecting Treatment and Control Villages
In each of the 63 municipalities selected by ASA for the project implementation, a Municipal Committee (MC) formed by local leaders selects between 6 and 12 villages that will enter a lottery. Typically, MCs select the neediest villages in the municipality – those with higher rates of poverty, lowest water availability, etc. Selected villages enter a lottery that randomly allocates three villages to control status and the rest into treatment. No households in control villages will receive P1+2 for the next two years, at least. The lottery is done electronically through a purpose-built website (http://impactocisternas.org) during a meeting organized by the member of the implementing agency, who was previously trained to use the website and explain the procedure to the members of the MC.
The completion of the 63 lotteries yielded a sample frame of 565 villages, split in control (N = 171) and treatment (N = 394) villages.
Eligibility of Households
To be eligible, households must meet the following criteria:
1) Be a beneficiary of the Primeira Agua program, which delivered a 16,000-liter cistern for human consumption.
2) Be enrolled in Cadastro Unico, a national registry of vulnerable households, maintained by MDS. Cadastro Unico is the main gateway for vulnerable households to enter many of the federal government’s social policies, such as Bolsa Família. It contains detailed information on household composition and demographics, income, dwelling characteristics, location, etc.
Selecting Eligible Households
1) MDS develops a list of all eligible households in the 63 municipalities, using data from the Cadastro Unico joined with the list of past beneficiaries of the Primeira Agua program.
2) The research team then randomly chooses and ranks 20 households and provides this list to the implementer, who must go down the list in that order until six households receive the treatment. The list has up to 20 household names to allow for substitution of households that cannot be reached (e.g., in cases of death or migration) due to outdated Cadunico records.
The research team can then apply steps 1 and 2 to select comparable households in control villages to be interviewed (see below for details on data collection), thus avoiding differential selection of households in control and treatment villages.
Selection of Individuals Within Households
To increase the ability to test for program effects on intra-household bargaining power, the research design introduces intra-household variation in the assignment of cisterns to the man or woman. Although the unit of analysis is a household, treatment is awarded to one individual within the household. This is the individual that receives training and signs documents under his or her name.
The research design introduces intra-household gender variation in the assignment of treatment using Cadastro Unico’s list of household members. The research team randomly allocates the program to the man or woman identified as the household head or his/her spouse. This works only as a nudge since implementers are free to enroll anyone in the household into treatment. Whether or not this is nudge is powerful enough to create intra-household variation, is a question that will have to wait until implementers finish enrolling households into the program. If powerful, it will provide additional variation in treatment assignment, which can be used to test the hypotheses that P1+2 attenuates intra-household gender imbalances, identifies factors that may prevent or promote effects on reducing gender inequality, and the trade-offs (if any) between the goals of improving climate adaptation and reducing gender imbalances.
Data collection:
This research project comprises two waves of primary data collection using household survey instruments developed by the research team. In each wave, 2,930 households from treatment and control villages will be interviewed. MDS held a public tender process to hire a specialized firm to administer the survey instruments. The first wave (baseline) took place between July and September of 2018. The second wave (endline) will reinterview the same households 18 months after the treatment is delivered.
There are two survey instruments. The first must be answered by the person who spends more time in rural activities within the household and contains questions on land use and crop choices; water usage; income; labor, farm production; technology use; household consumption; food security; and subjective well-being. The second must be answered by the female (head of household or spouse) and covers topics on female autonomy, social capital, and decision-making power.
Finally, we plan to use automated response mobile calls to collect high-frequency data on consumption patterns and food security. This is a cost-effective way to collect data on issues such as food consumption and reactions to weather shocks, which require high-frequency monitoring. Up to 85% of the target households use mobile phones, and cellphone coverage in the region is good.
Proposed data analysis:
To estimate the impact of the cisterns on outcome y (e.g., income, indexes of women empowerment, crop diversification, etc.), we will estimate the following population regression equation by ordinary least squares (OLS):
y_icmt=α_m+βT_icm+λy_(icmt-1)+ ϵ_icmt (1)
Where T_icmt takes the value of one if family i lives in a treatment village and zero otherwise, α_m is a municipality fixed-effect, θ_t is a time fixed-effect, and ϵ_icmt is the idiosyncratic error term. Standard errors will be clustered at the village level, c.
To test for effects of giving treatment to women vs. men on gender issues (bargaining power, autonomy, and resource allocation), we will estimate the following regression model:
y_icmt=κ_m+ϕT_icm+γT_icm×W_imct+ωy_(icmt-1)+ ξ_icmt (2)
where W_imct indicates that a woman was the individual assigned to receive treatment. In case this indicator does not follow the random assignment, we can still estimate intention-to-treat effects and local average treatment effects by applying an instrumental variables approach, using variation from the intra-household random assignment.